@@ -248,8 +248,8 @@ jobs:
248248 ctest --test-dir cmake-out -R test_cuda_allocator --output-on-failure -V
249249 ctest --test-dir cmake-out -R test_cuda_mutable_state --output-on-failure -V
250250
251- export -model-cuda-artifact :
252- name : export -model-cuda-artifact
251+ test -model-cuda-e2e :
252+ name : test -model-cuda-e2e-${{ matrix.model.name }}-${{ matrix.quant }}
253253 # Skip this job if the pull request is from a fork (HuggingFace secrets are not available).
254254 # Path-filtered on push: mirrors the workflow-level pull_request `paths:`
255255 # filter so push commits that don't touch CUDA-relevant paths skip
@@ -352,7 +352,7 @@ jobs:
352352 repo : " facebook"
353353 name : " dinov2-small-imagenet1k-1-layer"
354354 quant : " quantized-int4-weight-only"
355- # Qwen3 int4-weight-only: no downstream job consumes this artifact
355+ # Keep the existing Qwen3 coverage to non-quantized and tile-packed.
356356 - model :
357357 repo : " Qwen"
358358 name : " Qwen3-0.6B"
@@ -377,7 +377,7 @@ jobs:
377377 name : " whisper-large-v3-turbo"
378378 quant : " non-quantized"
379379 with :
380- timeout : 150
380+ timeout : 240
381381 secrets-env : EXECUTORCH_HF_TOKEN
382382 runner : ${{ (matrix.model.name == 'Qwen3.5-35B-A3B-HQQ-INT4' || matrix.model.name == 'gemma-4-31B-it-GGUF') && 'mt-l-x86iavx512-11-125-a100' || 'mt-l-x86aavx2-29-113-a10g' }}
383383 gpu-arch-type : cuda
@@ -419,153 +419,10 @@ jobs:
419419 echo "::endgroup::"
420420 fi
421421
422- source .ci/scripts/export_model_artifact.sh cuda "${{ matrix.model.repo }}/${{ matrix.model.name }}" "${{ matrix.quant }}" "${RUNNER_ARTIFACT_DIR}"
423-
424- test-model-cuda-e2e :
425- name : test-model-cuda-e2e
426- # Same path filter as export-model-cuda-artifact above. Also explicitly
427- # gated on the export job succeeding — when needs: jobs are *skipped*
428- # (e.g. fork PR), GitHub still evaluates this if:, so without the
429- # explicit success-check this job would run and then fail trying
430- # to download an artifact that was never produced.
431- needs : [changed-files, export-model-cuda-artifact, run-decision]
432- if : |
433- needs.export-model-cuda-artifact.result == 'success' &&
434- (
435- contains(needs.changed-files.outputs.changed-files, 'backends/cuda') ||
436- contains(needs.changed-files.outputs.changed-files, 'backends/aoti') ||
437- contains(needs.changed-files.outputs.changed-files, '.github/workflows/cuda.yml') ||
438- contains(needs.changed-files.outputs.changed-files, '.ci/scripts/test-cuda-build.sh') ||
439- contains(needs.changed-files.outputs.changed-files, '.ci/scripts/export_model_artifact.sh') ||
440- contains(needs.changed-files.outputs.changed-files, '.ci/scripts/test_model_e2e.sh') ||
441- needs.run-decision.outputs.is-full-run == 'true'
442- )
443- uses : pytorch/test-infra/.github/workflows/linux_job_v3.yml@main
444- permissions :
445- id-token : write
446- contents : read
447- strategy :
448- fail-fast : false
449- matrix :
450- model :
451- - repo : " mistralai"
452- name : " Voxtral-Mini-3B-2507"
453- - repo : " mistralai"
454- name : " Voxtral-Mini-4B-Realtime-2602"
455- - repo : " nvidia"
456- name : " diar_streaming_sortformer_4spk-v2"
457- - repo : " openai"
458- name : " whisper-small"
459- - repo : " openai"
460- name : " whisper-large-v3-turbo"
461- - repo : " google"
462- name : " gemma-3-4b-it"
463- - repo : " nvidia"
464- name : " parakeet-tdt"
465- - repo : " facebook"
466- name : " dinov2-small-imagenet1k-1-layer"
467- - repo : " SocialLocalMobile"
468- name : " Qwen3.5-35B-A3B-HQQ-INT4"
469- - repo : " unsloth"
470- name : " gemma-4-31B-it-GGUF"
471- quant :
472- - " non-quantized"
473- - " quantized-int4-tile-packed"
474- - " quantized-int4-weight-only"
475- exclude :
476- # TODO: enable int4-weight-only on gemma3.
477- - model :
478- repo : " google"
479- name : " gemma-3-4b-it"
480- quant : " quantized-int4-weight-only"
481- # Qwen3.5 MoE uses a prequantized checkpoint, only tile-packed
482- - model :
483- repo : " SocialLocalMobile"
484- name : " Qwen3.5-35B-A3B-HQQ-INT4"
485- quant : " non-quantized"
486- - model :
487- repo : " SocialLocalMobile"
488- name : " Qwen3.5-35B-A3B-HQQ-INT4"
489- quant : " quantized-int4-weight-only"
490- # Gemma 4 31B uses a prequantized checkpoint, only tile-packed
491- - model :
492- repo : " unsloth"
493- name : " gemma-4-31B-it-GGUF"
494- quant : " non-quantized"
495- - model :
496- repo : " unsloth"
497- name : " gemma-4-31B-it-GGUF"
498- quant : " quantized-int4-weight-only"
499- # Voxtral Realtime only supports int4-tile-packed on CUDA
500- - model :
501- repo : " mistralai"
502- name : " Voxtral-Mini-4B-Realtime-2602"
503- quant : " non-quantized"
504- - model :
505- repo : " mistralai"
506- name : " Voxtral-Mini-4B-Realtime-2602"
507- quant : " quantized-int4-weight-only"
508- # Sortformer currently supports only non-quantized export
509- - model :
510- repo : " nvidia"
511- name : " diar_streaming_sortformer_4spk-v2"
512- quant : " quantized-int4-tile-packed"
513- - model :
514- repo : " nvidia"
515- name : " diar_streaming_sortformer_4spk-v2"
516- quant : " quantized-int4-weight-only"
517- # DINOv2 currently supports only non-quantized export
518- - model :
519- repo : " facebook"
520- name : " dinov2-small-imagenet1k-1-layer"
521- quant : " quantized-int4-tile-packed"
522- - model :
523- repo : " facebook"
524- name : " dinov2-small-imagenet1k-1-layer"
525- quant : " quantized-int4-weight-only"
526- - model :
527- repo : " nvidia"
528- name : " parakeet-tdt"
529- quant : " quantized-int4-weight-only"
530- # Whisper: cover all quant types and both sizes with minimal combos
531- # whisper-small: non-quantized only
532- # whisper-large-v3-turbo: int4-tile-packed + int4-weight-only
533- - model :
534- repo : " openai"
535- name : " whisper-small"
536- quant : " quantized-int4-tile-packed"
537- - model :
538- repo : " openai"
539- name : " whisper-small"
540- quant : " quantized-int4-weight-only"
541- - model :
542- repo : " openai"
543- name : " whisper-large-v3-turbo"
544- quant : " non-quantized"
545- with :
546- timeout : 90
547- runner : ${{ (matrix.model.name == 'Qwen3.5-35B-A3B-HQQ-INT4' || matrix.model.name == 'gemma-4-31B-it-GGUF') && 'mt-l-x86iavx512-11-125-a100' || 'mt-l-x86aavx2-29-113-a10g' }}
548- gpu-arch-type : cuda
549- gpu-arch-version : " 13.0"
550- use-custom-docker-registry : false
551- submodules : recursive
552- download-artifact : ${{ matrix.model.repo }}-${{ matrix.model.name }}-cuda-${{ matrix.quant }}
553- ref : ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
554- script : |
555- # OSDC runners can't reach the public PyPI CDN that download.pytorch.org's
556- # transitive deps resolve to. Pre-install torch's pure-python deps from the
557- # in-cluster pypi-cache and drop the default cpu extra-index so the cuda
558- # torch wheel is the only candidate.
559- export PIP_EXTRA_INDEX_URL=
560- # fsspec is pinned to satisfy datasets' fsspec[http]<=2025.3.0 so the later
561- # examples install doesn't try to downgrade it from the public CDN.
562- pip install filelock typing-extensions "setuptools<82" sympy networkx jinja2 "fsspec[http]<=2025.3.0" numpy pillow
563- # OSDC mounts HF_HOME read-only at /mnt/hf_cache; redirect to a writable dir
564- # (RUNNER_TEMP, or /tmp when RUNNER_TEMP isn't writable inside the container).
565- export HF_HOME="${RUNNER_TEMP:-/tmp}/hf_cache"
566- mkdir -p "${HF_HOME}" 2>/dev/null || export HF_HOME=/tmp/hf_cache
567- mkdir -p "${HF_HOME}"
568- source .ci/scripts/test_model_e2e.sh cuda "${{ matrix.model.repo }}/${{ matrix.model.name }}" "${{ matrix.quant }}" "${RUNNER_ARTIFACT_DIR}"
422+ RUN_EXPORT=1 source .ci/scripts/test_model_e2e.sh cuda \
423+ "${{ matrix.model.repo }}/${{ matrix.model.name }}" \
424+ "${{ matrix.quant }}" \
425+ "${RUNNER_ARTIFACT_DIR}"
569426
570427 test-muse-glimmer-cuda-e2e :
571428 name : test-muse-glimmer-cuda-e2e-${{ matrix.variant }}-${{ matrix.mode }}
@@ -639,13 +496,12 @@ jobs:
639496
640497 test-cuda-pybind :
641498 name : test-cuda-pybind
642- # This job downloads models exported by export-model-cuda-artifact and runs them using pybind.
643- # Same gating as test-model-cuda-e2e — explicit success-check on the
644- # export job so a skipped export (fork PR, non-sampled push, no path
645- # match) auto-skips this job too.
646- needs : [changed-files, export-model-cuda-artifact, run-decision]
499+ # This job downloads models exported by test-model-cuda-e2e and runs them using pybind.
500+ # Explicitly check the producer job so a skipped run (fork PR,
501+ # non-sampled push, or no path match) auto-skips this job too.
502+ needs : [changed-files, test-model-cuda-e2e, run-decision]
647503 if : |
648- needs.export -model-cuda-artifact .result == 'success' &&
504+ needs.test -model-cuda-e2e .result == 'success' &&
649505 (
650506 contains(needs.changed-files.outputs.changed-files, 'backends/cuda') ||
651507 contains(needs.changed-files.outputs.changed-files, 'backends/aoti') ||
0 commit comments